What Is Artificial Intelligence?
What exactly is artificial intelligence? Although most attempts to define complex and widely used terms precisely are exercises in futility, it is useful to draw at least an approximate boundary around the concept to provide a perspective on the discussion that follows. To do this, we propose the following by no means universally accepted definition. Artificial intelligence (AI) is the study of how to make computers do things which, at the moment, people do better. This definition is, of course, somewhat ephemeral becauşe of its reference to the current state of computer science. And it fails to include some areas of potentially very large impact, namely problems that cannot now be solved well by either computers or people. But it provides a good outline of what constitutes artificial intelligence, and it avoids the philosophical issues that dominate attempts to define the meaning of either artificial or intelligence. Interestingly, though, it suggests a similarity with philosophy at the same time it is avoiding it. Philosophy has always been the study of those branches of knowledge that were so poorly understood that they had not yet become separate disciplines in their own right. As fields such as mathematics or physics became more advanced, they broke off from philosophy. Perhaps if AI succeeds it can reduce itself to the empty set.
What then are some of the problems contained within AI? Much of the early work in the field focused on formal tasks, such as game playing and theorem proving. Samuel wrote a checkers-playing program that not only played games with opponents but also used its experience at those games to improve its later performance. Chess also received a good deal of attention. The Logic Theorist was an early attempt to prove mathematical theorems. It was able to prove several theorems from the first chapter of Whitehead and Russell's Principia Mathematica. Gelernter's theorem prover explored another area of mathematics: geometry. Game playing and theorem proving share the property that people who do them well are considered to be displaying intelligence, Despite this, it appeared initially that computers could perform well at those tasks simply by being fast at exploring a large number of solution paths and then selecting the best one. It was thought that this process required very little knowledge and could therefore be programmed easily. As we will see later, this assumption tumed out to be false since no computer is fast enough to overcome the combinatorial explosion generated by most problems.
Another early foray into AI focused.on the sort of problem solving that we do every day when we decide how to get to work in the moming, often called commonsense reasoning. It includes reasoning about physical objects and their relationships to each other (e.g., an object can be in only one place at a time), as well as reasoning about actions and their consequences (e.g., if you let go of something, it will fall to the floor and maybe break). To investigate this sort of reasoning, Newell, Shaw, and Simon built the General Problem Solver (GPS), which they applied to several commonsense tasks as well as to the problem of performing symbolic manipulations of logical expressions. Again, no attempt was made to create a program with a large amount of knowledge about a particular problem domain. Only quite simple tasks were selected.
As AI research progressed and techniques for handling larger amounts of world knowledge were developed, some progress was made on the tasks just described and new tasks could reasonably be attempted. These include perception (vision and speech), natural language understanding, and problem solving in specialized domains such as medical diagnosis and chemical analysis.
Perception of the world around us is crucial to our survival. Animals with much less intelligence than people are capable of more sophisticated visual perception than are current machines. Perceptual tasks are difficult because they involve analog (rather than digital) signals; the signals are typically very noisy and usually a large number of things (some of which may be partially obscuring others) must be perceived at once. The problems of perception are discussed in greater detail in Chapter 21.
The ability to use language to communicate a wide variety of ideas is perhaps the most important thing that separates humans from the other animals. The problem of understanding spoken language is a perceptual problem and is hard to solve for the reasons just discussed. But suppose we simplify the problem by restricting it to written language. This problem, usually referred to as natural language understanding, is still extremely difficult. In order to understand sentences about a topic, it is necessary to know not only a lot about the language itself (its vocabulary and grammar) but also a good deal about the topic so that unstated assumptions can be recognized. We discuss this problem again later in this chapter and then in more detail in Chapter 15.
In addition to these mundane tasks, many people can also perform one or maybe more specialized tasks in which carefully acquired expertise is necessary. Examples of such tasks include engineering design, scientific discovery, medical diagnosis, and financial planning. Programs that can solve problems in these domains also fall under the aegis of artificial intelligence. Figure 1.1 lists some of the tasks that are the targets of work in AI.
A person who knows how to perform tasks from several of the categories shown in the figure learns the necessary skills in a standard order. First perceptual, linguistic, and commonsense skills are learned. Later (and of course for some people, never) expert skills such as engineering, medicine, or finance are acquired. It might seem to make sense then that the earlier skills are easier and thus more amenable to computerized duplication than are the later, more specialized ones. For this reason, much of the initial AI work was concentrated in those early areas. But it tums out that this naive assumption is not right. Although expert skills reyuire knowledge that many of us do not have, they
often require much less knowledge than do the more mundane skills and that knowledge is usually easier to represenl and deal with inside programs.
As a result, the problem areas where AI is now flourishing most as a practical discipline (as opposed to a purely research one) are primarily the domains that require only specialized expertise without the assistance ofcommonsense knowledge. There are now thousands of programs called expert system,s in day-to-day operation throughout all areas of industry and government. Each of these systems attempts to solve part, or perhaps all, of a practical, significant problem that previously required scarce human expertise. In Chapter 20 we examine several of these systems and explore techniques for constructing them.
Before embarking on a study of specific AI problems and solution techniques, it is important at least to discuss, if not to answer, the following four questions:
The next four sections of this chapter address these questions. Following that is a surveg ot some AI books that may be of interest and a summary of the chapter.
At the heart of researeh in artificial intelligence lies what Newell and Simon [1976] call the physical symbol system hypothesis. They define a physical symbol system as follows:
They then state the hypothesis as
This hypothesis is only a hypothesis. There appears to be no way to prove or disprove it on logical grounds. So it must be subjected to empirical validation. We may find that it is false. We may find that the bulk of the evidence says that it is true. But the only way to determine its truth is by experimentation.
Computers provide the perfect medium for this experimentation since they can be programmed to simulate any physical symbol system we like. This ability of computers to serve as arbitrary symbol manipulators was noticed very early in the history of computing. Lady Lovelace made the following observation about Babbage's proposed Analytical Engine in 1842:
As it has become increasingly easy to build computing machines, so it has become inereasingly possible to conduct empirical investigations of the physical symbol system hypothesis. In each such investigation, a particular task that might be regarded as requiring intelligence is selected. A program to perform the task is proposed and then tested. Although we have not been completely successful at creating programs that parform all the selected tasks, most scientists believe that many of the problems that have been encountered will ultimately prove to be surmountable by more sophisticated programs than we have yet produced.
Evidence in support of the physical symbol system hypothesis has come not only from areas such as game playing, where one might most expect to find it, but also from areas such as visual perception, where it is more tempting to suspect the influence of subsymbolic processes. However, subsymbolic models (for example, neural networks) are beginning to challenge symbolic ones at such low-level tasks. Such models arc discussed in Chapter 18. Whether certain subsymbolic models conflict with the physical symbol system hypothesis is a topic still under debate (e.g., Smolensky [1988]). And it is important to note that even the success of subsymbolic systems is not necessarily evidence against the hypothesis. It is often possible to accomplish a task in more than one way.
One interesting attempt to reduce a patticularly human activity, the understanding ofjokes, to a process of symbol manipulation is provided in the book Mathematics and Humor [Paulos,1980]. It is, of course, possible that the hypothesis will turn out to be only partially true. Perhaps physical symbol systems will prove able to model some aspects of human intelligence and not others. Only time and effort will tell.
The importance of the physical symbol system hypothesis is twofold. It is a significant theory of the nature of human intelligence and so is ofgreat interest to psychologists. It also forms the basis of the belief that it is possible to build programs that can perform intelligent tasks now performed by people. Our major concern here is with the latter of these implications, although as we will soon see, the two issues are not unrelated.
Artificial intelligence problems span a very broad spectrum. They appear to have very little in common except that they are hard. Are there any techniques that are appropriate for the solution of a variety of these problems? The answer to this question is yes, there are. What, then, if anything, can we say about those techniques besides the fact that they manipulate symbols? How could we tell if those techniques might be useful in solving other problems, perhaps ones not traditionally regarded as AI tasks? The rest of this book is an attempt to answer those questions in detail. But before we begin examining closely the individual techniques, it is enlightening to take a broad look at them to see what properties they ought to possess.
One of the few hard and fast results to come out of the first three decades of AI research is that intelligence requires knowledge. To compensate for its one overpowering asset, indispensability knowledge possesses some less desirable properties, including:
So where does this leave us in our attempt to define AI techniques? We are forced to conclude that an AI technique is a method that exploits knowledge that should be represented in such a way that:
Although AI techniques must be designed in keeping with these constraints imposed by AI problems, there is some degree of independence between problems and problemsolving techniyues. ft is possible to solve AI problems without using AI techniques (although, as we suggested above, those solutions are not likely to be very good). And it is possible to apply AI techniques to the solution of non-AI problems. This is likely to be a good thing to do for problems that possess many of the same characteristics as do AI problems. In order to try to characterize AI techniques in as problem-independent a way as possible, let's look at two very different problems and a series of approaches for solving each of them.
In this section, we present a series of three programs to play tic-tac-toe. The programs in this series increase in:
Thus they move toward being representations of what we call AI techniques.
Data Structures
An element contains the value 0 if the corresponding square is blank, 1 if it is filled with an X, or 2 if it is filled with an O.
The Algorithm
To make a move, do the following:
Comments
This program is very efficient in terms of time. And, in theory, it could play an optimal game of tic-tac-toe. But it has several disadvantages:
The technique embodied in this program does not appear to meet any ofour requirements for a good AI technique. Let's see if we can do better.
Data Structures
The Algorithm
The main algorithm uses three subprocedures:
The algorithm has a built-in strategy for each move it may have to make. It makes the odd-numbered moves if it is playing X, the even-numbered moves if it is playing O. The strategy for each turn is as follows:
Turn=1 Go(1) (upper left corner).
Turn=2 If Board[5] is blank, Go(5), else Go(1).
Turn=3 If Board[9] is blank, Go(9), else Go(3).
Turn=4 If Posswin(X) is not 0, then Go(Posswin(X)) [i.e., block opponent's win], else Go(Make2).
Turn=5 If Posswin(X) is not 0 then Go(Posswin(X)) [i.e., win] else if Posswin(O) is not 0, then Go(Posswin(O)) [i.e., block win], else if Board[7] is blank, then Go(7), else Go(3). [Here the program is trying to make a fork.]
Turn=6 If Posswin(O) is not 0 then Go(Posswin(O)), else if Posswin(X) is not 0, then Go(Posswin(X)), else Go(Make2).
Turn=7 If Posswin(X) is not 0 then Go(Posswin(X)), else if Posswin(O) is not 0, then Go(Posswin(O)), else go anywhere that is blank.
Turn=8 If Posswin(O) is not 0 then Go(Posswin(O)), else if Posswin(X) is not 0, then Go(Posswin(X)), else go anywhere that is blank.
Turn=9 Same as Turn=7.
Comments
This program is not quite as efficient in terms of time as the first one since it has to check several conditions before making each move. But it is a lot more efficient in terms of space. It is also a lot easier to understand the program's strategy or to change the strategy if desired. But the total strategy has still been figured out in advance by the programmer. Any bugs in the programmer's tic-tac-toe playing skill will show up in the program's play. And we still cannot generalize any of the program's knowledge to a different domain, such as three-dimensional tic-tac-toe.
This program is identical to Program 2 except for one change in the representation of the board. We again represent the board as a nine-element vector, but this time we assign board positions to vector elements as follows:
Notice that this numbering of the board produces a magic square: all the rows, columns, and diagonals sum to 15. This means that we can simplify the process of checking for a possible win. In addition to marking the board as moves are made, we keep a list, for each player, of the squares in which he or she has played. To check for a possible win for one player, we consider each pair of syuares owned by that player and compute the difference between 15 and the sum of the two squares. If this difference is not positive or if it is greater than 9, then the original two squares were not collinear and so can be ignored. Otherwise, if the square representing the difference is blank, a move there will produce a win. Since no player can have more than four squares at a time, there will be many fewer squares examined using this scheme than there were using the more straightforward approach of Program 2. This shows how the choice of representation can have a major impact on the efficiency of a problem-solving program.
Comments
This comparison raises an interesting question about the relationship between the way people solve problems and the way computers do. Why do people find the row-scan approach easier while the number-counting approach is more efficient for a computer? We do not know enough about how people work to answer that question completely. One part of the answer is that people are parallel processors and can look at several parts of the board at once, whereas the conventional computer must look at the squares one at a time. Sometimes an investigation of how people solve problems sheds great light on how computers should do so. At other times, the differences in the hardware of the two seem so great that different strategies seem best. As we learn more about problem solving both by people and by machines, we may know better whether the same representations and algorithms are best for both people and machines. We will discuss this question further in Section 1.4.
Data Structures
The Algorithm
To decide on the next move, look ahead at the board positions that result from each possible move. Decide which position is best (as described below), make the move that leads to that position, and assign the rating of that best move to the current position.
To decide which of a set of board positions is best, do the following for each of them :
This algorithm will look ahead at various sequences of moves in order to find a sequence that leads to a win. It attempts to maximize the likelihood of winning, while assuming that the opponent will try to minimize that likelihood. This algorithm is called the minimax procedure, and it is discussed in detail in Chapter 12.
Comments
This program will require much more time than either of the others since it must search a tree representing all possible move sequences before making each move. But it is superior to the other programs in one very big way: It could be extended to handle games more complicated than tic-tac-toe, for which the exhaustive enumeration approach of the other programs would completely fall apart. It can also be augmented by a variety of specific kinds of knowledge about games and how to play them. For example, instead of considering all possible next moves, it might consider only a subset of them that are determined, by some simple algorithm, to be reasonable. And, instead of following each series of moves until one player wins, it could search for a limited time and evaluate the merit of each resulting board position using some static function.
Program 3 is an example of the use of an AI technique. For very small problems, it is less efficient than a variety of more direct methods. However, it can be used in situations where those methods would fail.
In this section we look at a series of programs that read in English text and then answer questions, also stated in English, about that text. This task differs from the last one in that it is more difficult now to state formally and precisely what our problem is and what constitutes correct solutions to it. For example, suppose that the input text were just the single sentence
Russia massed troops on the Czech border.
Then either of the following question-answering dialogues might occur (and in fact did occur with the POLITICS program [Carbonell,1980)):
Q: Why did Russia do this?
A: Because Russia thought that it could take political cuntrol of Czechoslovakia by sending troops.
Q: What should the United States do?
A: The United States should intervene militarily.
Q: Why did Russia do this`?
A: Because Russia wanted to increase its political inftuence over Czechoslovakia.
Q: What should the United States do?
A: The United States should denounce the Russian action in the United Nations.
In the POLITICS program, answers were constructed by considering both the input text and a separate model ofthe beliefs and actions ofvarious political entities, including Russia. When the model is changed, as it was between these two dialogues, the system's answers also change. In this example, the first dialogue was produced when POLITICS was given a model that was intended to correspond to the beliefs of a typical American conservative (circa 1977). The second dialogue occurred when POLITICS was given a model that was intended to correspond to the beliefs of a typical American liberal (of the same vintage).
The general point here is that defining what it means to produce a correct answer to a question may be very hard. Usually, question-answering programs define what it means to be an answer by the procedure that is used to compute the answer. Then their authors appeal to other people to agree that the answers found by the program "make sense" and so to confirm the model of question answering defined in the program. This is not completely satisfac;tory, but no better way of defining the problem has yet been found. For lack of a better method, we will do the same here and illustrate three definitions of question answering, each with a corresponding program that implements the definition.
In order to be able to compare the three programs, we illustrate all of them using the following text:
We will also attempt to answer each of the following questions with each program
Q1: What did Mary go shopping for?
Q2: What did Mary find that she liked?
Q3: Did Mary buy anything?
This program attempts to answer questions using the literal input text. It simply matches text fragments in the questions against the input text.
Data Structures
The Algorithm
To answer a question, do the following:
Examples
Q1: The template "What did x y" matches this question and generates the text pattern "Mary go shopping for z. ' After the pattern-substitution step, this pattern is expanded to a set of patterns including "Mary goes shopping for z," and "Mary went shopping for z. ' The Iatter pattern matches the input text; the program, using a convention that variables match the longest possible string up to a sentence delimiter (such as a period), assigns z the value, "a new coat," which is given as the answer.
Q2: Unless the template set is very large, allowing for the insertion of the object of "find" between it and the modifying phrase "that she liked," the insertion of the word "really" in the text, and the substitution of "she" for "Mary," this question is not answerable. If all of these variations are accounted for and the question can be answered, then the response is "a red one."
Q3: Since no answer to this question is contained in the text, no answer will be found.
Comments
This approach is clearly inadequate to answer the kinds of yuestions people could answer after reading a simple text. Even its ability to answer the most direct questions is delicately dependent on the exact form in which questions are stated and on the variations that were anticipated in the design of the templates and the pattern substitutions that the system uses. In fact, the sheer inadequacy of this program to perform the task may make you wonder how such an approach could even be proposed. This program is substantially farther away from being useful than was the initial program we looked at for tic-tac-toe. Is this just a strawman designed to make some other technique look good in comparison'? In a way, yes, but it is worth mentioning that lhe approach that this program uses, namely matching patterns, performing simple text substitutions, and then forming answers using straightforward combinations of canned text and sentence fragments located by the matcher, is the same approach that is used in one of the most famous "AI" programs ever written-ELIZA, which we discuss in Section 6.4.1. But, as you read the rest of this sequence of programs, it should become clear that what we mean by the term "artificial intelligence" does not inelude programs such as this except by a substantial stretching of definitions.
This program first converts the input text into a structured internal form that attempts to capture the meaning of the sentences. It also converts questions into that form. It finds answers by matching structured forms against each other.
Data Structures
of representation is that entities in the representation derive their meaning from their connections to other entities. In the figure, only the entities defined by the sentence are shown. But other entities, corresponding to concepts that the program knew about before it read this sentence, also exist in the representation and can be referred to within these new structures. In this example, for instance, we refer to the entities Mary, Coat (the general concept of a coat of which Thing1 is a specific instance), Liking (the general concept of liking), and Finding (the general concept of finding).
The Algorithm
Convert the InputText into structured form using the knowledge contained in EnglishKnow. This may require considering several different potential structures, for a variety ofreasons, including the fact that English words can be ambiguous, English grammatical structures can be ambiguous, and pronouns may have several possible antecedents.
Then, to answer a question, do the following:
Examples
Q1: This question is answered straightforwardly with, "a new coat."
Q2: This one also is answered successfully with, "a red coat."
Q3: This one, though, cannot be answered, since there is no direct response to it in the text.
Comments
This approach is substantially more meaning (knowledge)-based than that of the first program and so is more effective. It can answer most questions to which replies are contained in the text, and it is much less brittle than the first program with respect to the exact forms of the text and the questions. As we expect, based on our experience with the pattern recognition and tic-tac-toe programs, the price we pay for this increased ftexibility is time spent searching the various knowledge bases (i.e., EnglishKnow, StructuredText).
One word of warning is appropriate here. The problem of producing a knowledge base for English that is powerful enough to handle a wide range of English inputs is very difficult. It is discussed at greater length in Chapter 15. In addition, it is now recognized that knowledge of English alone is not adequate in general to enable a program to build the kind of structured representation shown here. Additional knowledge about the world with which the text deals is often required to support lexical and syntactic disambiguation and the correct assignment of antecedents to pronouns, among other things. For example, in the text
it is not possible to determine what the word "she" refers to without knowledge about the roles of customers and salespeople in stores. To see this, contrast the correct antecedent of "she" in that text with the correct antecedent for the first occurrence of "she" in the following example:
In the simple case illustrated in our coat-buying example, it is possible to derive correct answers to our first two questions without any additional knowledge about stores or coats, and the fact that some such additional information may be necessary to support question answering has already been illustrated by the failure of this program to find an answer to question 3. Thus we see that although extracting a structured representation of the meaning of the input text is an improvement over the meaning-free approach of Program I, it is by no means sufficient in general. So we need to look at an even more sophisticated (i.e., knowledge-rich) approach, which is what we do next.
This program converts the input text into a structured form that contains the meanings of the sentences in the text, and then it combines that form with other structured forms that describe prior knowledge about the objects and situations involved in the text. It answers questions using this augmented knowledge structure.
Data Structures
The Algorithm
Convert the InputText into structured form using both the knowledge contained in EnglishKnow and that contained in WorldModel. The number of possible structures will usually be greater now than it was in Program 2 because so much more knowledge is being used. Sometimes, though, it may be possible to consider fewer possibilities by using the additional knowledge to filter the alternatives.
To answer a question, do the following:
Examples
Q1: Same as Program 2.
Q2: Same as Program 2.
Q3: Now this question can be answered. The shopping script is instantiated for this text, and because of the last sentence, the path through step 14 of the script is the one that is used in forming the representation of this text. When the script is instantiated M' is bound to the structure representing the red coat (because the script says that M' is what gets taken home and the text says that a red coat is what got taken home). After the script has been instantiated, IntegratedText contains several events that are taken from the script but that are not described in the original text, including the event "Mary buys a red coat" (from step 10 of the script). Thus, using the integrated text as the basis for question answering allows the program to respond "She bought a red coat."
Comments
This program is more powerful than either of the first two because it exploits more knowledge. Thus it, like the final program in each of the other two sequences we have examined, is exploiting what we call AI techniques. But, again, a few caveats are in order. Even the techniques we have exploited in this program are not adequate for complete English question answering. The most important thing that is missing from this program is a general reasoning (inference) mechanism to be used when the requested answer is not contained explicitly even in IntegratedText, but that answer does follow logically from the knowledge that is there. For example, given the text
it should be possible to answer the question
with the reply
But to do so requires knowing that one cannot be two places at once and then using that fact to conclude that Mary could not have been home because she was shopping instead. Thus, although we avoided the inference problem temporarily by building IntegratedText, which had some obvious inferences built into it, we cannot avoid it forever. It is simply not practical to anticipate all legitimate inferences. In later chapters, we look at ways of providing a general inference mechanism that could be used to support a program such as the last one in this series.
This limitation does not contradict the main point of this example though. In fact, it is additional evidence for that point, namely, an effective question-answering procedure must be one based soundly on knowledge a:id the computational use of that knowledge. The purpose of AI techniques is to support this effective use of knowledge.
We have just examined two series of programs to solve two very different problems. In each series, the final program exemplifies what we mean by an AI technique. These two programs are slower to execute than the earlier ones in their respective series, but they illustrate three important AI techniques:
For the solution of hard problems, programs that exploit these techniques have several advantages over those that do not. They are much less fragile; they will not be thrown off completely by a small perturbation in their input. People can easily understand what the program's knowledge is. And these techniques can work for large problems where more direct methods break down.
We have still not given a precise definition of an AI technique. It is probably not possible to do so. But we have given some examples of what one is and what one is not. Throughout the rest of this book, we talk in great detail about what one is. The definition should then become a bit clearer, or less necessary.
Before we set out to do something, it is a good idea to decide exactly what we are trying to do. So we must ask ourselves, "What is our goal in trying to produce programs that do the intelligentthings that people do?" Are we trying to produce programs that do the tasks tbe same way people do? Or, are we attempting to produce programs that simply do the tasks in whatever way appears easiest? There have been AI projects motivated by each of these goals.
Efforts to build programs that perform tasks the way people do can be divided into two classes. Programs in the first class attempt to solve problems that do not really fit our definition of an AI task. They are problems that a computer could easily solve, although that easy solution would exploit mechanisms that do not seem to be available to people. A classical example of this class of program is the Elementary Perceiver and Memorizer (EPAM) [Feigenbaum, 1963], which memorized associated pairs of nonsense syllables. Memorizing pairs of nonsense syllables is easy for a computer. Simply input them. To retrieve a response syllable given its associated stimulus one, the computer just scans for the stimulus syllable and responds with the one stored next to it. But this task is hard for people. EPAM simulated one way people might perform the task. It built a discrimination net through which it could find images of the syllables it had seen. It also stored, with each stimulus image, a cue that it could later pass through the discrimination net to try to find the correct response image. But it stored as a cue only as much information about the response syllable as was necessary to avoid ambiguity at the time the association was stored. This might be just the first letter, for example. But, of course, as the discrimination net grew and more syllables were added, an old cue might no longer be sufficient to identify a response syllable uniquely. Thus EPAM, like people, sometimes "forgot" previously learned responses. Many people regard programs in this first class to be uninteresting, and to some extent they are probably right. These programs can, however, be useful tools for psychologists who want to test theories of human performance.
The second class of programs that attempt to model human performance are those that do things that fall more clearly within our definition of AI tasks; they do things that are not trivial for the computer. There are several reasons one might want to model human performance at these sorts of tasks:
This last motivation is probably the most pervasive of the four. It motivated several very early systems that attempted to produce intelligent behavior by imitating people at the level of individual neurons. For examples of this, see the early theoretical work of McCulloch and Pitts [1943], the work on perceptrons, originally developed by Frank Rosenblatt but best described in Perceptrons [Minsky and Papert,1969] and Design for a Brain [Ashby, 1952]. It proved impossible, however, to produce even minimally intelligent behavior with such simple devices. One reason was that there were severe theoretical limitations to the particular neural net architecture that was being used. More recently, several new neural net architectures have been proposed. These structures are not subject to the same theoretical limitations as were perceptrons. These new architectures are loosely called connectionist, and they have been used as a basis for several learning and problem-solving programs. We have more to say about them in Chapter 18. Also, we must consider that while human brains are highly parallel devices, most current computing systems are essentially serial engines. A highly successful parallel technique may be computationally intractable on a serial computer. But recently, partly because of the existence of the new family of parallel cognitive models, as well as because of the general promise of parallel computing, there is now substantial interest in the design of massively parallel machines to support AI programs.
Human cognitive theories have also influenced AI to look for higher-level (i.e., far above the neuron level) theories that do not require massive parallelism for their implementation. An early example of this approach can be seen in GPS, which are discussed in more detail in Section 3.6. This same approach can also be seen in much current work in natural language understanding. The failure of straightforward syntactic parsing mechanisms to make much of a dent in the problem of interpreting English sentences has led many people who are interested in natural language understanding by machine to look seriously for inspiration at what little we know about how people interpret language. And when people who are trying to build programs to analyze pictures discover that a filter function they have developed is very similar to what we think people use, they take heart that perhaps they are on the right track.
As you can see, this last motivation pervades a great many areas of AI research. In fact, it, in conjunction with the other motivations we mentioned, tends to make the distinction between the goal of simulating human performance and the goal of building an intelligent program any way we can seem much less different than they at first appeared. In either case, what we really need is a good model of the processes involved in intelligentreasoning. The field ofcognitive science, in which psychologists, linguists, and computer scientists all work together, has as its goal the discovery of such a model. For a good survey of the variety of approaches contained within the field, see Norman [1981], Anderson [1985], and Gardner [1985].
One of the most important questions to answer in any scientific or engineering research project is "How will we know if we have succeeded?" Artificial intelligence is no exception. How will we know if we have constructed a machine that is intelligent? That question is at least as hard as the unanswerable question "What is intelligence?" But can we do anything to measure our progress?
In 1950, Alan Turing proposed the following method for determining whether a machine can think. His method has since become known as the Turing test. To conduct this test, we need two people and the machine to be evaluated. One person plays the role of the interrogator, who is in a separate room from the computer and the other person. The interrogator can ask questions of either the person or the computer by typing questions and receiving typed responses. However, the interrogator knows them only as A and B and aims to determine which is the person and which is the machine. The goal of the machine is to fool the interrogator into believing that it is the person. If the machine succeeds at this, then we will conclude that the machine can think. The machine is allowed to do whatever it can to fool the interrogator. So, for example, if asked the question "How much is 12,324 times 73,981?" it could wait several minutes and then respond with the wrong answer [Turing,1963].
The more serious issue, though, is the amount of knowledge that a machine would need to pass the Turing test. Turing gives the following example of the sort of dialogue a machine would have to be capable of:
Interrogator: 1n the first line of your sonnet which reads "Shall I compare thee to a summer's day," would not "a spring day" do as well or better?
A: It wouldn't scan.
Interrogator: How about "a winter's day." That would scan all right.
A: Yes,but nobody wants to be compared to a winter's day.
Interrogator: Would you say Mr.Pickwick reminded you of Christmas?
A: In a way.
Interrogator: Yet Christmas is a winter's day,and I do not think Mr.Pickwick would mind the comparison.
A: I don't think you're serious.By a winter's day one means a typical winter's day,rather than a special one like Christmas.
It will be a long time before a computer passes the Turing test. Some people believe none ever will. But suppose we are willing to settle for less than a complete imitation of a person. Can we measure the achievement of AI in more restricted domains?
Often the answer to this question is yes. Sometimes it is possible to get a fairly precise measure of the achievement of a program. For example, a program can acquire a chess rating in the same way as a human player. The rating is based on the ratings of players whom the program can beat. Already programs have acquired chess ratings higher than the vast majority of human players. For other problem domains, a less precise measure of a program's achievement is possible. For example, DENDRAL is a program that analyzes organic compounds to determine their structure. It is hard to get a precise measure of DENDRAL's level of achievement compared to human chemists, but it has produced analyses that have been published as original research results. Thus it is certainly performing competently.
In other technical domains, it is possible to compare the time it takes for a program to complete a task to the time required by a person to do the same thing. For example, there are several programs in use by computer companies to configure particular systems to customers' needs (of which the pioneer was a program called R1). These programs typically require minutes to perform tasks that previously required hours of a skilled engineer's time. Such programs are usually evaluated by looking at the bottom linewhether they save (or make) money.
For many everyday tasks, though, it may be even harder to measure a program's performance. Suppose, for example, we ask a program to paraphrase a newspaper story. For problems such as this, the best test is usually just whether the program responded in a way that a person could have.
If our goal in writing a program is to simulate human performance at a task, then the measure of success is the extent to which the program's behavior corresponds to that performance, as measured by various kinds of experiments and protocol analyses. In this we do not simply want a program that does as well as possible. We want one that fails when people do. Various techniques developed by psychologists for comparing individuals and for testing models can be used to do this analysis.
We are forced to conclude that the question of whether a machine has intelligence or can think is too nebulous to answer precisely. But it is often possible to construct a computer program that meets some performance standard for a particular task. That does not mean that the program does the task in the best possible way. It means only that we understand at least one way of doing at least part of a tdsk. When we set out to design an AI program, we should attempt to specify as well as possible the criteria for success for that particular program functioning in its restricted domain. For the moment, that is the best we can do.
There are a great many sources of information about artificial intelligence. First, some survey books: The broadest are the multi-volume Handbook of Artificial Intelligence [Barr et al., 1981] and Encyclopedia of Artificial Intelligence [Shapiro and Eckroth, 1987], both of which contain articles on each of the major topics in the field. Four other books that provide good overviews of the field are Artificial Intelligence [Winston, 1984], Introduction to Artificial Intelligence [Charniak and McDermott,1985], Logical Foundations of Artificial intelligence [Genesereth and Nilsson, 1987], and The Elements of Artificial Intelligence [Tanimoto, 1987]. Of more restricted scope is Principles of Artificial Intelligence [Nilsson, 1980], which contains a formal treatment of some general-purpose AI techniques.
The history of research in artificial intelligence is a fascinating story, related by Pamela McCorduck [1979] in her book Machines Who Think. Because almost all of what we call AI has been developed over the last 30 years, McCorduck was able to conduct her research for the book by actually interviewing almost all of the people whose work was influential in forming the field.
Most of the work conducted in AI has been originally reported in journal articles, conference proceedings, or technical reports. But some of the most interesting of these papers have later appeared in special collections published as books. Computers and Thought [Feigenbaum and Feldman, 1963] is a very early collection of this sort. Later ones include Simon and Siklossy [1972], Schank and Colby [1973], Bobrow and Collins [1975], Waterman and Hayes-Roth [1978], Findler [1979], Webber and Nilsson [1981], Halpem [1986], Shrobe [1988], and several others that are mentioned in later chapters in connection with specific topics.
The major journal of AI research is called simply Artificial intelligence. In addition, Cognitive Science is devoted to papers dealing with the overlapping areas of psychology linguistics, and artificial intelligence. AI Magazine is a more ephemeral, less technical magazine that is published by the Ameriean Association for Artificial Intelligence (AAAI). IEEE Expert and several other journals publish papers about expert systems in a wide variety of application domains.
Since 1969, there has been a major AI conference, the International Joint Conference on Artificial Intelligence (IJCAI), held every two years. The proceedings of these conferences give a good picture of the work that was taking place at the time. The other important AI conference, held three out of every four years starting in 1980, is sponsored by the AAAI, and its proceedings, too, are published.
In addition to these general references, there exists a whole array of papers and books describing individual AI projects. Rather than trying to list them all here, they are referred to as appropriate throughout the rest of this book.
What conclusions can we draw from this hurried introduction to the major questions of AI? The problems are varied, interesting, and hard. If we solve them, we will have useful programs and perhaps a better understanding of human thought. We should do the best we can to set criteria so that we can tell if we have solved the problems, and then we must try to do so.
How actually to go about solving these problems is the topic for the rest of this book. We need methods to help us solve AI's serious dilemma:
Our goal in AI is to construct working programs that solve the problems we are interested in. Throughout most of this book we focus on the design of representation mechanisms and algorithms that can be used by programs to solve the problems. We do not spend much time discussing the programming process required to turn these designs into working programs. In theory, it does not matier how this process is carried out, in what language it is done, or on what machine the product is run. In practice, of course, it is often much easier to produce a program using one system rather than another. Specifically, AI programs are easiest to build using languages that have been designed to support symbolic rather than primarily numeric computation.
For a variety of reasons, LISP has historically been the most commonly used language for AI programming. We say little explicitly about LISP in this book, although we occasionally rely on it as a notation. There used to be several competing dialects of LISP, but Common Lisp is now accepted as a standard. If you are unfamiliar with LISP consult any of the following sources: LISP [Winston and Horn, 1989], Common Lisp [Hennessey,1989], Common LISPcraft [Wilensky,1986], and Common Lisp: A Gentle Introduction to Symbolic Computation [Touretzky,1989a]. For a complete description of Common Lisp, see Common Lisp: The Reference [Steele,1990]. Another language that is often used for AI programming is PROLOG, which is described briefly in Chapter 6. And increasingly, as AI makes its way into the conventional programming world, AI systems are being written in general purpose programming languages such as C. One reason for this is that AI programs are ceasing to be standalone systems; instead, they are becoming components of larger systems, which may include conventional programs and databases of various forms. Real code does not form a big part of this book precisely because it is possible to implement the techniques we discuss in any of several languages and it is important not to confuse the ideas with their specific implementations. But you should keep in mind as you read the rest of this book that both the knowledge structures and the problem-solving strategies we discuss must ultimately be coded and integrated into a working program.
AI is still a young discipline. We have learned many things, some of which are presented in this book. But it is still hard to know exactly the perspective from which those things should be viewed. We cannot resist quoting an observation made by Lady Lovelace more than 100 years ago:
She was talking about Babbage's Analytical Engine. But she could have been describing artificial intelligence.